Unbounded Differentially Private Quantile and Maximum Estimation
Authors: David Durfee
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically compare our unbounded algorithm with the state-of-the-art algorithms in the bounded setting. For inner quantiles, we find that our method often performs better on non-synthetic datasets. For the maximal quantiles, which we apply to differentially private sum computation, we find that our method performs significantly better. In Section 5, we test our method compared to the previous techniques on synthetic and real world datasets. |
| Researcher Affiliation | Industry | David Durfee Anonym Inc. david@anonymco.com |
| Pseudocode | Yes | Algorithm 1 Above Threshold. Algorithm 2 Unbounded quantile mechanism. Algorithm 3 Iterative Exponential Mechanism. Algorithm 4 Fully unbounded quantile mechanism. |
| Open Source Code | Yes | We provide the code in the appendix for easier reproducibility. |
| Open Datasets | Yes | We borrow the same setup and datasets as Gillenwater et al. (2021); Kaplan et al. (2022). Two datasets will come from Soumik (2019) with 11,123 data points... Two datasets will come from Dua and Graf (2019) with 48,842 data points... Soumik (2019). Goodreads-books dataset. In https://www.kaggle.com/jealousleopard/ goodreadsbooks. Dua, D. and Graf, C. (2019). Uci machine learning repository. In http://archive.ics.uci.edu/ ml. |
| Dataset Splits | No | No explicit training/validation/test splits (e.g., percentages, sample counts, or references to predefined splits) are provided. The paper mentions randomly sampling 1000 datapoints and iterating 100 times for evaluation, which is a resampling strategy rather than a defined split for reproducibility. |
| Hardware Specification | No | No specific hardware (e.g., GPU/CPU models, memory, or cloud instance types) used for running the experiments is mentioned in the paper. |
| Software Dependencies | No | The paper mentions using "the in-built quantile function in the numpy library" and "exponential noise from the numpy library" but does not specify version numbers for numpy or Python. |
| Experiment Setup | Yes | We set \\u03b5= 1 as in the previous works, which will require setting \\u03b51 = \\u03b52 = 1/2 for the call to Above Threshold in our method. For our method we set \\u03b2= 1.001 for all datasets. For our method we only use q= 0.99. For the others we use the best performance for q {0.95, 0.96, 0.97, 0.98, 0.99}. |